Patient questions between visits need a responsible recipient

In a fictional hospital outpatient service, a patient sends a message after an appointment. They have received another letter and cannot tell whether it changes what was agreed. An AI assistant explains the wording clearly and suggests contacting the clinic. The conversation looks helpful. Yet the clinic may still have no assigned task, no person expecting to respond and no evidence that the uncertainty was resolved.

That is the decision I would put to hospital leadership before introducing an AI companion between visits: which patient questions will become a responsibility the service can actually fulfil? The ability to continue a conversation creates possibilities. Its value depends on what happens when the next useful step requires a person, a record or an appointment.

A useful question from the public conversation

In a Stanford Online episode published on 25 November 2025, Syed Mohiuddin, Matt Lungren and Justin Norden discuss AI support outside consultations. Their conversation prompted this article's question about the work between visits. It offers a future-facing perspective, rather than evidence that an autonomous service improves patient outcomes. The full episode is available here. [1]

I would turn that ambition into a narrower design question. Can a patient see what the service has understood, which team has accepted responsibility and what will happen next? These are distinct states. A message can be delivered without being assessed. A question can receive an accurate explanation while still needing a decision that the explanation cannot provide.

Two interventions with different evidence

A 2024 UC San Diego study randomized 52 volunteer primary-care physicians to immediate or delayed access to drafted replies. Reading took longer; time composing replies did not change significantly. This short, single-system evaluation did not measure whether patients’ concerns were resolved. [2]

For leadership, I would keep the measurement question open beyond drafting. Time spent preparing a reply describes one part of the service. My proposed evaluation would also ask whether the patient reached the right person and understood the next step. A product comparison based on writing time alone cannot answer those additional questions.

The 2025 PRO-TECT report compared weekly symptom reporting with care-team alerts against usual care among patients with metastatic cancer in 52 randomized US oncology practices. Emergency visits decreased; survival did not differ significantly. This combined cancer-care intervention provides no direct evidence about generative chatbots or general hospital messaging. [3]

My inference is to evaluate the route from a patient’s concern to an appropriate response as a whole. For a local trial, I would record which components were actually available: a reply, a person able to act and a route back if the concern remains. That would let leadership see where a promising conversation still leaves an unfinished task.

Specify the service boundary before the interface

For an initial deployment, I would choose one existing outpatient pathway and a defined set of questions. Locating an approved document, explaining the organisation of a visit and asking for an individual clinical decision involve different work. A message can contain all three. A label such as “administrative enquiry” should therefore be revisable when the actual content requires another route.

The patient-facing explanation should describe the available service, its staffed hours and what happens outside its scope. Those statements must match actual capacity. If a channel is not continuously monitored, the interface must not imply continuous clinical observation. The organisation must retain its established route for urgent concerns; this article proposes no clinical triage thresholds.

There is relevant existing guidance. The US SAFER communication guide recommends recording acknowledgement, escalating unread messages and providing cover when a recipient is unavailable. It also addresses clear patient expectations about response times. These are safety recommendations for electronic communication, rather than evidence of AI effectiveness or a statement of German legal requirements. [4]

Follow one concern through the handover

For the fictional letter question, I would distinguish receipt, acceptance by the responsible team, an agreed action and closure. Each status should have a plain meaning. An automated receipt confirms arrival. Acceptance means that the designated service has taken the task. Closure records how the concern was addressed. An unresolved concern keeps its status, responsible recipient and next action.

The handover should retain the patient's original question alongside any AI summary. It should indicate what the assistant already said and which uncertainty remains. Otherwise, staff may respond to the summary while overlooking the reason the patient asked. The receiving professional must be able to correct the interpretation and return the case to the appropriate route.

A transfer also needs a fallback when nobody accepts it. Leadership should assign the service responsible for that gap, including absences and handovers between shifts. Merely copying more people into a message leaves the question of acceptance open. The system should make an unaccepted task visible to someone authorised to resolve it.

I would test these states with invented conversations before inviting patients: a question containing two requests, an absent recipient, an unclear reply and a patient who cannot use the portal. The test asks whether the organisation knows what happens next. Passing this rehearsal supplies no evidence of clinical safety or successful operation at scale.

Measure access to resolution

For a bounded evaluation, I would count concerns requiring follow-up and record how many acquired an accountable recipient. I would distinguish time to receipt, time to professional assessment and time to an agreed next step. The relevant time limits need local professional agreement. A rapid automated acknowledgement should remain visible as its own event.

Closure also needs checking from the patient's side. In a sample, ask whether the person understands what will happen and whom to contact if it does not. A read receipt alone cannot answer that question. Unanswered follow-up, repeated contact and a question reopened after apparent closure should remain in the evaluation.

Additional questions may represent previously unmet need becoming visible. They may also reflect confusing explanations or duplicated channels. Neither interpretation follows from message volume alone. Examine the reasons alongside staff effort across the whole service. Include people who declined digital access or needed another channel, so the evaluation does not describe only those who could complete the digital interaction.

This matters to me as a physician and founder building aiomics: a technical capability becomes an organisational commitment when people rely on it. I would approve a between-visit service around a defined scope, a staffed response route and evidence that patients reach an appropriate next step. The next investment decision should follow those observations, including the questions the service still cannot resolve.